Unsupervised Detection of Epileptic Seizures from EEG Signals: A Channel-Specific Analysis of Long-Term Recordings

Abstract

As long-term EEG recordings are becoming more and more frequent in clinical practice, the volume of EEG data that require annotation by clinical experts is growing exponentially, highlighting the need for automated epileptic seizure detection systems. Supervised methodologies based on machine learning techniques have been proposed as a solution, but they require annotated EEG data from multiple patients for training, while unsupervised methodologies, which do not have such limitations, are scarce in the literature. Thus, an unsupervised seizure detection methodology is developed to offer high seizure detection performance with substantial reduction in the time and effort required to inspect large volumes of EEG signals. The ictal rhythmical activity is detected by analyzing the spectral information of each EEG channel independently in order to find intensive accumulation of signal energy over the fundamental delta, theta and alpha frequency bands. The ictal rhythmical activity that is expressed during a seizure is detected using a set of four simple seizure detection conditions. The proposed methodology is validated using the public CHB-MIT EEG database, and the results suggest that an average sensitivity of 95.1% can be obtained with a false detection rate of 10.13 FP/h, while 95% of the each patient's EEG recordings is automatically rejected as non-ictal.